Claude finds 11-year-old remote flaw in Linux TCP/IP stack

What's going on

The Linux kernel's networking code contains a critical security flaw that affects nearly all Linux distributions, including Debian, Red Hat, Suse and Ubuntu. Anthropic's AI Claude discovered it while systematically testing major open-source projects. According to heise, the bug lets attackers inject and run their own code over the network (remote code execution). Its severity is CVSS 9.8, the highest danger level. The kernel team has already released a fix, and the first distributors have published updates.

The technical details

The bug is a race condition. The outcome depends on the timing of several processes, in this case two CPU cores. An attacker can trick the victim's kernel into having one CPU use memory that another CPU has just freed (a so-called use-after-free). This leads to unintended memory accesses. The attacker can coordinate them so that code they have injected gets executed.

According to heise, Claude didn't just suspect the flaw. It proved it with a working exploit. Anthropic then reported the bug to the Linux kernel team.

Eleven years undetected

The faulty code was checked in on October 2, 2015, as part of kernel version 4.4. Several experts reviewed and approved it. Only a systematic search by an AI brought the bug to light. On October 15, kernel maintainer Greg Kroah-Hartman confirmed that CVE-2026-98432 is fixed in kernel versions 6.12.113, 6.18.56, 7.2.10 and 7.3-rc7.

All IPv4 implementations since kernel 4.4 are affected. The IPv6 implementation is not.

What admins should do now

According to heise, distributors are working flat out on updated kernels for all affected products that are still supported. The advisories from Debian, Red Hat, Suse and Ubuntu show the current status. Administrators should install the patched kernels as soon as they become available.

How dangerous is it really?

Little is known so far. The only known exploit comes from Anthropic and is not public. It targets Multipath TCP (MPTCP), a TCP option that a server service has to request explicitly. According to heise, only Go applications enable it by default. It's unclear whether an attack requires an MPTCP service to be running on the target system. There are no known real-world attacks so far.

Context

The heise author sees the case as a particularly striking example of how AI systems can uncover security flaws in code that has grown over decades. The Linux kernel's networking code is considered especially well maintained, and other software is usually in worse shape. Until now, such flaws were protected by being hard to find and even harder to exploit. If LLMs (large language models) largely automate that search, a big wave of vulnerabilities is coming. The author talks about accumulated "security debt" that is now coming due with compound interest. That is an assessment, not a measurement. It remains open how easily the flaw can be exploited without MPTCP, and whether attackers already knew about it.